DATA ANALYSIS STEPS

The detailed steps implemented for processing the EEG data are explained below in a stepwise manner. The EEG data obtained from the experimental results have been analyzed using Brain Vision Analyzer (version 2.1).

Changing Sampling Rate: The initial sampling rate while setting up the

Cognionics system in experimental stage is 500 Hz. An even frequency resolution can be achieved by having a sampling frequency that is a power of 2, i.e., 512 or 256 Hz instead 500 Hz (Lin et al., 2007). So, to obtain more fine-grained resolution, downsampling (number of samples per second have been decreased) to 256 HZ has been performed applying spline interpolation (see Figure 5.2).

Optimizing EEG Channel Selection: Generally, multichannel EEG is used in BCI; performing the channel selection enhances the signal processing accuracy by discarding irrelevant channels and promoting relevant channels (Arvaneh et al., 2011). In this study, we have 5 channels which do not contribute to neural activity analysis and these channels are eliminated. Figure 5.3. depicts the channel optimization process. Such customized approach can lead to best signal processing and classification accuracy.

Figure 5.3. Optimizing Channel Selection

Raw Data Inspection/Artifact Rejection: Raw Data Inspection is used for marking artifacts like body movements, environmental noise, eye blinks, and eye movements; it is also sensitive to large offset voltages. These marked data portions are considered as “bad

intervals” and are rejected by ocular correction ICA (Independent Component Analysis) based on the rejection criteria (Plank, 2013). As an initial step, an automatic raw data inspection was applied using a built-in algorithm of the Brain Vision Analyzer in a semi- automatic mode at the individual channel level. This algorithm excludes intervals of 200 ms if the voltage of an activity exceeds 50 μV/ms or if it is less than 0.5 μV for a time frame of 100 ms (Figure 5.4 – 5.7) (Ulrich & Hewig, 2014). Implementing this technique in a semi-automatic mode not only helps to discard artifacts but also aids to inspect the data manually (Beste et al., 2015). Figure 5.4 explains the inspection mode as semi- automatic for raw data inspection process. Figure 5.5 and 5.6 provides information related to maximum and minimum voltage criteria. Figure 5.7 is the EEG signal obtained after performing raw data inspection.

Figure 5.5. Raw Data Inspection: Maximum Voltage Criteria_1

Figure 5.6. Raw Data Inspection: Minimum Voltage Criteria_2

Ocular Correction ICA: Among the EEG artifacts, ocular or eye movements and eye blinks are considered the most common and notorious artifacts (Minas et al., 2014). These ocular artifacts pose serious problems for interpretation and analysis of EEG signals and can be removed using Ocular Correction ICA.

The regression-based method is the most popular among all Ocular Artifact (OA) removal approaches in the time-frequency domain (Croft & Barry, 2000a; Croft & Barry, 2000b). Regression-based methods can reduce ocular artifacts very effectively if they employ ocular EOG (Electrooculography measuring eye movement) channels (Li et al., 2006). Ocular Correction ICA is not completely dependent on ocular channels (as is, in turn, the regression-based Ocular Correction); the ICA algorithm is not fully reliant on ocular channels and delivers robust components for vertical and horizontal eye

movements with scalp channels as well (Plank, 2013). Therefore, in absence of dedicated ocular channels, Brain Vision Analyzer recommends using scalp channels that report respective artifacts adequately. For detecting and rejecting vertical movement (VEOG), AFF5h has been considered as a common reference (see Figure 5.9). Generally, it is recommended to perform Ocular Correction ICA in semi-automatic mode and assess carefully and confirm the selected components as ocular artifacts. Figure 5.8 represents the mode selection for the ocular correction ICA. Figure 5.9 represents the reference channel selection for the and Figure 5.10 represents process of identifying and removing eye-blinks while performing ocular correction ICA.

Figure 5.8. Ocular Correction ICA: Mode Selection

Figure 5.9. Ocular Correction ICA: Reference Channel Selection

Filtering: Applying digital filters is considered a common approach to reject EEG epochs containing artifacts with certain pre-selected voltage threshold. However, the amount of data may become unacceptable when muscle movements and blinks occur too frequently in some subjects (Small, 1971; Li and Principe, 2006). To filter the selected voltage, Infinite Impulse Response (IIR) filters are applied (Sanei &

Chambers, 2007; Wang et al., 2011) (see Figure 5.11). IIR filters are digital filters used in digital signal processing applications. As shown in Figure 5.11, the recorded EEG signals were analog bandpass filters between 0.1 Hz (Low Pass Filter) and 100 Hz (High Pass Filter); additionally, notch filter was applied at 60 Hz to substantially remove external noise related to line power frequencies.

Segmentation: EEG data can be divided into interval-based epochs to perform further analysis (Bender et al., 2004; Nickel et al., 2006). The processed data is

segmented into four divisions based on retrospective process tracing in the experiment stage representing the time frames of experienced user states. As, shown in Figures 5.12 and 5.13, the corrected data has been used to set the newly segmented data manually with respective start and end timestamps (resting, boredom, flow, anxiety); each of the 30 second EEG epochs have been further divided into 100 equal segments and were averaged to obtain enhanced accuracy in results. Figure 5.12 represents the manual division options for segmentation and Figure 5.13 represents the specific timeframes of each user’s state.

Figure 5.13. EEG Signal Segmentation: Time Frames of User’s States

Fast Fourier Transformation (FFT): The segmented EEG signals are in time- domain (i.e., time on the x-axis); to perform spectral band analysis, these EEG signals need to be converted into frequency-domain (i.e., frequency on x-axis) (Wang et al., 2011). FFT decomposes the time domain signals into frequency domain. By using a built-in algorithm in Brain Vision Analyzer, FFT has been applied to transform the time-domain EEG epochs into equivalent frequency-domain epochs. As shown in figure 5.14, the FFT values of theta, alpha, and mid-beta for resting, boredom, flow, and anxiety were extracted using FFT band export option provided by Brain Vision

Analyzer. The mean values of EEG power in different frequency bands (theta, alpha, and mid-beta) and at different brain regions (frontal, temporal, parietal, and occipital) were calculated to identify the neural correlates of the flow state (Kubota et al., 2001). Electrodes in the left frontal region (i.e., AFF1 (F1), AFF3 (F3), and AFF5 (F5)), left parietal region (i.e., CPP1h(P1), CPP3h(P3), CPP5h(P5), CPP7h(P7)), and occipital

region (i.e.., O1h(O1), Oz(Oz), O2h(O2)) were pooled to form a cluster. Finally, paired t- tests were performed to assess the hypotheses.

Figure 5.14. Exporting FFT Values for Theta Spectral Band